Dermato: A Deep Learning based Application for Acne Subtype and Severity Detection

Abstract

The article introduces a dermatology application named “Dermato”. It is an assistive application tool for dermatologists to quickly assess and classify acne lesions into respective types. The application is built using HTML, CSS, JavaScript, Flask, and MySQL database. The image processing and segmentation have been implemented using various deep learning modules like TensorFlow, OpenCV, and variations of CNNs. After training these models on customized datasets, the results were quite unreliable and the low accuracy made the model overfit or underfit the data. But once hyperparameter tuning was performed, the model performance improved and the classification results were mostly accurate. Another challenge faced was maintaining the authenticity of the dataset. This was achieved under the guidance of field professionals such as dermatologists. Even though the existing systems were able to accurately classify the facial regions into respective skin disorders (like acne, pimples, scars, moles, and dark circles, etc.), there exists no particular system which can further classify acne into its subtypes such as acne comedo, acne papules, acne pustules, acne scars, etc. A few recent techniques offering solutions for the above-mentioned problem use CNN-based models trained on humongous datasets such as DermNet and CelebA along with enhanced deep learning methods such as image segmentation. Dermato builds upon these techniques and focuses on further sub-classification of these acne regions into various subtypes of acne. The system uses a customized dataset, different pre-processing techniques, and segmentation strategies to deliver reliable results.